fs/ext4/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  linux/fs/ext4/file.c
   4  *
   5  * Copyright (C) 1992, 1993, 1994, 1995
   6  * Remy Card (card@masi.ibp.fr)
   7  * Laboratoire MASI - Institut Blaise Pascal
   8  * Universite Pierre et Marie Curie (Paris VI)
   9  *
  10  *  from
  11  *
  12  *  linux/fs/minix/file.c
  13  *
  14  *  Copyright (C) 1991, 1992  Linus Torvalds
  15  *
  16  *  ext4 fs regular file handling primitives
  17  *
  18  *  64-bit file support on 64-bit platforms by Jakub Jelinek
  19  *      (jj@sunsite.ms.mff.cuni.cz)
  20  */
  21
  22 #include <linux/time.h>
  23 #include <linux/fs.h>
  24 #include <linux/iomap.h>
  25 #include <linux/mount.h>
  26 #include <linux/path.h>
  27 #include <linux/dax.h>
  28 #include <linux/quotaops.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/uio.h>
  31 #include <linux/mman.h>
  32 #include <linux/backing-dev.h>
  33 #include "ext4.h"
  34 #include "ext4_jbd2.h"
  35 #include "xattr.h"
  36 #include "acl.h"
  37 #include "truncate.h"
  38
  39 static bool ext4_dio_supported(struct inode *inode)
  40 {
  41         if (IS_ENABLED(CONFIG_FS_ENCRYPTION) && IS_ENCRYPTED(inode))
  42                 return false;
  43         if (fsverity_active(inode))
  44                 return false;
  45         if (ext4_should_journal_data(inode))
  46                 return false;
  47         if (ext4_has_inline_data(inode))
  48                 return false;
  49         return true;
  50 }
  51
  52 static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
  53 {
  54         ssize_t ret;
  55         struct inode *inode = file_inode(iocb->ki_filp);
  56
  57         if (iocb->ki_flags & IOCB_NOWAIT) {
  58                 if (!inode_trylock_shared(inode))
  59                         return -EAGAIN;
  60         } else {
  61                 inode_lock_shared(inode);
  62         }
  63
  64         if (!ext4_dio_supported(inode)) {
  65                 inode_unlock_shared(inode);
  66                 /*
  67                  * Fallback to buffered I/O if the operation being performed on
  68                  * the inode is not supported by direct I/O. The IOCB_DIRECT
  69                  * flag needs to be cleared here in order to ensure that the
  70                  * direct I/O path within generic_file_read_iter() is not
  71                  * taken.
  72                  */
  73                 iocb->ki_flags &= ~IOCB_DIRECT;
  74                 return generic_file_read_iter(iocb, to);
  75         }
  76
  77         ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL,
  78                            is_sync_kiocb(iocb));
  79         inode_unlock_shared(inode);
  80
  81         file_accessed(iocb->ki_filp);
  82         return ret;
  83 }
  84
  85 #ifdef CONFIG_FS_DAX
  86 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
  87 {
  88         struct inode *inode = file_inode(iocb->ki_filp);
  89         ssize_t ret;
  90
  91         if (iocb->ki_flags & IOCB_NOWAIT) {
  92                 if (!inode_trylock_shared(inode))
  93                         return -EAGAIN;
  94         } else {
  95                 inode_lock_shared(inode);
  96         }
  97         /*
  98          * Recheck under inode lock - at this point we are sure it cannot
  99          * change anymore
 100          */
 101         if (!IS_DAX(inode)) {
 102                 inode_unlock_shared(inode);
 103                 /* Fallback to buffered IO in case we cannot support DAX */
 104                 return generic_file_read_iter(iocb, to);
 105         }
 106         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
 107         inode_unlock_shared(inode);
 108
 109         file_accessed(iocb->ki_filp);
 110         return ret;
 111 }
 112 #endif
 113
 114 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
 115 {
 116         struct inode *inode = file_inode(iocb->ki_filp);
 117
 118         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 119                 return -EIO;
 120
 121         if (!iov_iter_count(to))
 122                 return 0; /* skip atime */
 123
 124 #ifdef CONFIG_FS_DAX
 125         if (IS_DAX(inode))
 126                 return ext4_dax_read_iter(iocb, to);
 127 #endif
 128         if (iocb->ki_flags & IOCB_DIRECT)
 129                 return ext4_dio_read_iter(iocb, to);
 130
 131         return generic_file_read_iter(iocb, to);
 132 }
 133
 134 /*
 135  * Called when an inode is released. Note that this is different
 136  * from ext4_file_open: open gets called at every open, but release
 137  * gets called only when /all/ the files are closed.
 138  */
 139 static int ext4_release_file(struct inode *inode, struct file *filp)
 140 {
 141         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
 142                 ext4_alloc_da_blocks(inode);
 143                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
 144         }
 145         /* if we are the last writer on the inode, drop the block reservation */
 146         if ((filp->f_mode & FMODE_WRITE) &&
 147                         (atomic_read(&inode->i_writecount) == 1) &&
 148                         !EXT4_I(inode)->i_reserved_data_blocks)
 149         {
 150                 down_write(&EXT4_I(inode)->i_data_sem);
 151                 ext4_discard_preallocations(inode);
 152                 up_write(&EXT4_I(inode)->i_data_sem);
 153         }
 154         if (is_dx(inode) && filp->private_data)
 155                 ext4_htree_free_dir_info(filp->private_data);
 156
 157         return 0;
 158 }
 159
 160 /*
 161  * This tests whether the IO in question is block-aligned or not.
 162  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
 163  * are converted to written only after the IO is complete.  Until they are
 164  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
 165  * it needs to zero out portions of the start and/or end block.  If 2 AIO
 166  * threads are at work on the same unwritten block, they must be synchronized
 167  * or one thread will zero the other's data, causing corruption.
 168  */
 169 static bool
 170 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
 171 {
 172         struct super_block *sb = inode->i_sb;
 173         unsigned long blockmask = sb->s_blocksize - 1;
 174
 175         if ((pos | iov_iter_alignment(from)) & blockmask)
 176                 return true;
 177
 178         return false;
 179 }
 180
 181 static bool
 182 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
 183 {
 184         if (offset + len > i_size_read(inode) ||
 185             offset + len > EXT4_I(inode)->i_disksize)
 186                 return true;
 187         return false;
 188 }
 189
 190 /* Is IO overwriting allocated and initialized blocks? */
 191 static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
 192 {
 193         struct ext4_map_blocks map;
 194         unsigned int blkbits = inode->i_blkbits;
 195         int err, blklen;
 196
 197         if (pos + len > i_size_read(inode))
 198                 return false;
 199
 200         map.m_lblk = pos >> blkbits;
 201         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
 202         blklen = map.m_len;
 203
 204         err = ext4_map_blocks(NULL, inode, &map, 0);
 205         /*
 206          * 'err==len' means that all of the blocks have been preallocated,
 207          * regardless of whether they have been initialized or not. To exclude
 208          * unwritten extents, we need to check m_flags.
 209          */
 210         return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
 211 }
 212
 213 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
 214                                          struct iov_iter *from)
 215 {
 216         struct inode *inode = file_inode(iocb->ki_filp);
 217         ssize_t ret;
 218
 219         if (unlikely(IS_IMMUTABLE(inode)))
 220                 return -EPERM;
 221
 222         ret = generic_write_checks(iocb, from);
 223         if (ret <= 0)
 224                 return ret;
 225
 226         /*
 227          * If we have encountered a bitmap-format file, the size limit
 228          * is smaller than s_maxbytes, which is for extent-mapped files.
 229          */
 230         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
 231                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 232
 233                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
 234                         return -EFBIG;
 235                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
 236         }
 237
 238         return iov_iter_count(from);
 239 }
 240
 241 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
 242 {
 243         ssize_t ret, count;
 244
 245         count = ext4_generic_write_checks(iocb, from);
 246         if (count <= 0)
 247                 return count;
 248
 249         ret = file_modified(iocb->ki_filp);
 250         if (ret)
 251                 return ret;
 252         return count;
 253 }
 254
 255 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
 256                                         struct iov_iter *from)
 257 {
 258         ssize_t ret;
 259         struct inode *inode = file_inode(iocb->ki_filp);
 260
 261         if (iocb->ki_flags & IOCB_NOWAIT)
 262                 return -EOPNOTSUPP;
 263
 264         inode_lock(inode);
 265         ret = ext4_write_checks(iocb, from);
 266         if (ret <= 0)
 267                 goto out;
 268
 269         current->backing_dev_info = inode_to_bdi(inode);
 270         ret = generic_perform_write(iocb->ki_filp, from, iocb->ki_pos);
 271         current->backing_dev_info = NULL;
 272
 273 out:
 274         inode_unlock(inode);
 275         if (likely(ret > 0)) {
 276                 iocb->ki_pos += ret;
 277                 ret = generic_write_sync(iocb, ret);
 278         }
 279
 280         return ret;
 281 }
 282
 283 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
 284                                            ssize_t written, size_t count)
 285 {
 286         handle_t *handle;
 287         bool truncate = false;
 288         u8 blkbits = inode->i_blkbits;
 289         ext4_lblk_t written_blk, end_blk;
 290         int ret;
 291
 292         /*
 293          * Note that EXT4_I(inode)->i_disksize can get extended up to
 294          * inode->i_size while the I/O was running due to writeback of delalloc
 295          * blocks. But, the code in ext4_iomap_alloc() is careful to use
 296          * zeroed/unwritten extents if this is possible; thus we won't leave
 297          * uninitialized blocks in a file even if we didn't succeed in writing
 298          * as much as we intended.
 299          */
 300         WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
 301         if (offset + count <= EXT4_I(inode)->i_disksize) {
 302                 /*
 303                  * We need to ensure that the inode is removed from the orphan
 304                  * list if it has been added prematurely, due to writeback of
 305                  * delalloc blocks.
 306                  */
 307                 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
 308                         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 309
 310                         if (IS_ERR(handle)) {
 311                                 ext4_orphan_del(NULL, inode);
 312                                 return PTR_ERR(handle);
 313                         }
 314
 315                         ext4_orphan_del(handle, inode);
 316                         ext4_journal_stop(handle);
 317                 }
 318
 319                 return written;
 320         }
 321
 322         if (written < 0)
 323                 goto truncate;
 324
 325         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 326         if (IS_ERR(handle)) {
 327                 written = PTR_ERR(handle);
 328                 goto truncate;
 329         }
 330
 331         if (ext4_update_inode_size(inode, offset + written)) {
 332                 ret = ext4_mark_inode_dirty(handle, inode);
 333                 if (unlikely(ret)) {
 334                         written = ret;
 335                         ext4_journal_stop(handle);
 336                         goto truncate;
 337                 }
 338         }
 339
 340         /*
 341          * We may need to truncate allocated but not written blocks beyond EOF.
 342          */
 343         written_blk = ALIGN(offset + written, 1 << blkbits);
 344         end_blk = ALIGN(offset + count, 1 << blkbits);
 345         if (written_blk < end_blk && ext4_can_truncate(inode))
 346                 truncate = true;
 347
 348         /*
 349          * Remove the inode from the orphan list if it has been extended and
 350          * everything went OK.
 351          */
 352         if (!truncate && inode->i_nlink)
 353                 ext4_orphan_del(handle, inode);
 354         ext4_journal_stop(handle);
 355
 356         if (truncate) {
 357 truncate:
 358                 ext4_truncate_failed_write(inode);
 359                 /*
 360                  * If the truncate operation failed early, then the inode may
 361                  * still be on the orphan list. In that case, we need to try
 362                  * remove the inode from the in-memory linked list.
 363                  */
 364                 if (inode->i_nlink)
 365                         ext4_orphan_del(NULL, inode);
 366         }
 367
 368         return written;
 369 }
 370
 371 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
 372                                  int error, unsigned int flags)
 373 {
 374         loff_t offset = iocb->ki_pos;
 375         struct inode *inode = file_inode(iocb->ki_filp);
 376
 377         if (error)
 378                 return error;
 379
 380         if (size && flags & IOMAP_DIO_UNWRITTEN)
 381                 return ext4_convert_unwritten_extents(NULL, inode,
 382                                                       offset, size);
 383
 384         return 0;
 385 }
 386
 387 static const struct iomap_dio_ops ext4_dio_write_ops = {
 388         .end_io = ext4_dio_write_end_io,
 389 };
 390
 391 /*
 392  * The intention here is to start with shared lock acquired then see if any
 393  * condition requires an exclusive inode lock. If yes, then we restart the
 394  * whole operation by releasing the shared lock and acquiring exclusive lock.
 395  *
 396  * - For unaligned_io we never take shared lock as it may cause data corruption
 397  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
 398  *
 399  * - For extending writes case we don't take the shared lock, since it requires
 400  *   updating inode i_disksize and/or orphan handling with exclusive lock.
 401  *
 402  * - shared locking will only be true mostly with overwrites. Otherwise we will
 403  *   switch to exclusive i_rwsem lock.
 404  */
 405 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
 406                                      bool *ilock_shared, bool *extend)
 407 {
 408         struct file *file = iocb->ki_filp;
 409         struct inode *inode = file_inode(file);
 410         loff_t offset;
 411         size_t count;
 412         ssize_t ret;
 413
 414 restart:
 415         ret = ext4_generic_write_checks(iocb, from);
 416         if (ret <= 0)
 417                 goto out;
 418
 419         offset = iocb->ki_pos;
 420         count = ret;
 421         if (ext4_extending_io(inode, offset, count))
 422                 *extend = true;
 423         /*
 424          * Determine whether the IO operation will overwrite allocated
 425          * and initialized blocks.
 426          * We need exclusive i_rwsem for changing security info
 427          * in file_modified().
 428          */
 429         if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
 430              !ext4_overwrite_io(inode, offset, count))) {
 431                 inode_unlock_shared(inode);
 432                 *ilock_shared = false;
 433                 inode_lock(inode);
 434                 goto restart;
 435         }
 436
 437         ret = file_modified(file);
 438         if (ret < 0)
 439                 goto out;
 440
 441         return count;
 442 out:
 443         if (*ilock_shared)
 444                 inode_unlock_shared(inode);
 445         else
 446                 inode_unlock(inode);
 447         return ret;
 448 }
 449
 450 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
 451 {
 452         ssize_t ret;
 453         handle_t *handle;
 454         struct inode *inode = file_inode(iocb->ki_filp);
 455         loff_t offset = iocb->ki_pos;
 456         size_t count = iov_iter_count(from);
 457         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
 458         bool extend = false, unaligned_io = false;
 459         bool ilock_shared = true;
 460
 461         /*
 462          * We initially start with shared inode lock unless it is
 463          * unaligned IO which needs exclusive lock anyways.
 464          */
 465         if (ext4_unaligned_io(inode, from, offset)) {
 466                 unaligned_io = true;
 467                 ilock_shared = false;
 468         }
 469         /*
 470          * Quick check here without any i_rwsem lock to see if it is extending
 471          * IO. A more reliable check is done in ext4_dio_write_checks() with
 472          * proper locking in place.
 473          */
 474         if (offset + count > i_size_read(inode))
 475                 ilock_shared = false;
 476
 477         if (iocb->ki_flags & IOCB_NOWAIT) {
 478                 if (ilock_shared) {
 479                         if (!inode_trylock_shared(inode))
 480                                 return -EAGAIN;
 481                 } else {
 482                         if (!inode_trylock(inode))
 483                                 return -EAGAIN;
 484                 }
 485         } else {
 486                 if (ilock_shared)
 487                         inode_lock_shared(inode);
 488                 else
 489                         inode_lock(inode);
 490         }
 491
 492         /* Fallback to buffered I/O if the inode does not support direct I/O. */
 493         if (!ext4_dio_supported(inode)) {
 494                 if (ilock_shared)
 495                         inode_unlock_shared(inode);
 496                 else
 497                         inode_unlock(inode);
 498                 return ext4_buffered_write_iter(iocb, from);
 499         }
 500
 501         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend);
 502         if (ret <= 0)
 503                 return ret;
 504
 505         /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
 506         if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
 507                 ret = -EAGAIN;
 508                 goto out;
 509         }
 510
 511         offset = iocb->ki_pos;
 512         count = ret;
 513
 514         /*
 515          * Unaligned direct IO must be serialized among each other as zeroing
 516          * of partial blocks of two competing unaligned IOs can result in data
 517          * corruption.
 518          *
 519          * So we make sure we don't allow any unaligned IO in flight.
 520          * For IOs where we need not wait (like unaligned non-AIO DIO),
 521          * below inode_dio_wait() may anyway become a no-op, since we start
 522          * with exclusive lock.
 523          */
 524         if (unaligned_io)
 525                 inode_dio_wait(inode);
 526
 527         if (extend) {
 528                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 529                 if (IS_ERR(handle)) {
 530                         ret = PTR_ERR(handle);
 531                         goto out;
 532                 }
 533
 534                 ret = ext4_orphan_add(handle, inode);
 535                 if (ret) {
 536                         ext4_journal_stop(handle);
 537                         goto out;
 538                 }
 539
 540                 ext4_journal_stop(handle);
 541         }
 542
 543         if (ilock_shared)
 544                 iomap_ops = &ext4_iomap_overwrite_ops;
 545         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
 546                            is_sync_kiocb(iocb) || unaligned_io || extend);
 547         if (ret == -ENOTBLK)
 548                 ret = 0;
 549
 550         if (extend)
 551                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 552
 553 out:
 554         if (ilock_shared)
 555                 inode_unlock_shared(inode);
 556         else
 557                 inode_unlock(inode);
 558
 559         if (ret >= 0 && iov_iter_count(from)) {
 560                 ssize_t err;
 561                 loff_t endbyte;
 562
 563                 offset = iocb->ki_pos;
 564                 err = ext4_buffered_write_iter(iocb, from);
 565                 if (err < 0)
 566                         return err;
 567
 568                 /*
 569                  * We need to ensure that the pages within the page cache for
 570                  * the range covered by this I/O are written to disk and
 571                  * invalidated. This is in attempt to preserve the expected
 572                  * direct I/O semantics in the case we fallback to buffered I/O
 573                  * to complete off the I/O request.
 574                  */
 575                 ret += err;
 576                 endbyte = offset + err - 1;
 577                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
 578                                                    offset, endbyte);
 579                 if (!err)
 580                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
 581                                                  offset >> PAGE_SHIFT,
 582                                                  endbyte >> PAGE_SHIFT);
 583         }
 584
 585         return ret;
 586 }
 587
 588 #ifdef CONFIG_FS_DAX
 589 static ssize_t
 590 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
 591 {
 592         ssize_t ret;
 593         size_t count;
 594         loff_t offset;
 595         handle_t *handle;
 596         bool extend = false;
 597         struct inode *inode = file_inode(iocb->ki_filp);
 598
 599         if (iocb->ki_flags & IOCB_NOWAIT) {
 600                 if (!inode_trylock(inode))
 601                         return -EAGAIN;
 602         } else {
 603                 inode_lock(inode);
 604         }
 605
 606         ret = ext4_write_checks(iocb, from);
 607         if (ret <= 0)
 608                 goto out;
 609
 610         offset = iocb->ki_pos;
 611         count = iov_iter_count(from);
 612
 613         if (offset + count > EXT4_I(inode)->i_disksize) {
 614                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
 615                 if (IS_ERR(handle)) {
 616                         ret = PTR_ERR(handle);
 617                         goto out;
 618                 }
 619
 620                 ret = ext4_orphan_add(handle, inode);
 621                 if (ret) {
 622                         ext4_journal_stop(handle);
 623                         goto out;
 624                 }
 625
 626                 extend = true;
 627                 ext4_journal_stop(handle);
 628         }
 629
 630         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
 631
 632         if (extend)
 633                 ret = ext4_handle_inode_extension(inode, offset, ret, count);
 634 out:
 635         inode_unlock(inode);
 636         if (ret > 0)
 637                 ret = generic_write_sync(iocb, ret);
 638         return ret;
 639 }
 640 #endif
 641
 642 static ssize_t
 643 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
 644 {
 645         struct inode *inode = file_inode(iocb->ki_filp);
 646
 647         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 648                 return -EIO;
 649
 650 #ifdef CONFIG_FS_DAX
 651         if (IS_DAX(inode))
 652                 return ext4_dax_write_iter(iocb, from);
 653 #endif
 654         if (iocb->ki_flags & IOCB_DIRECT)
 655                 return ext4_dio_write_iter(iocb, from);
 656
 657         return ext4_buffered_write_iter(iocb, from);
 658 }
 659
 660 #ifdef CONFIG_FS_DAX
 661 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
 662                 enum page_entry_size pe_size)
 663 {
 664         int error = 0;
 665         vm_fault_t result;
 666         int retries = 0;
 667         handle_t *handle = NULL;
 668         struct inode *inode = file_inode(vmf->vma->vm_file);
 669         struct super_block *sb = inode->i_sb;
 670
 671         /*
 672          * We have to distinguish real writes from writes which will result in a
 673          * COW page; COW writes should *not* poke the journal (the file will not
 674          * be changed). Doing so would cause unintended failures when mounted
 675          * read-only.
 676          *
 677          * We check for VM_SHARED rather than vmf->cow_page since the latter is
 678          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
 679          * other sizes, dax_iomap_fault will handle splitting / fallback so that
 680          * we eventually come back with a COW page.
 681          */
 682         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
 683                 (vmf->vma->vm_flags & VM_SHARED);
 684         pfn_t pfn;
 685
 686         if (write) {
 687                 sb_start_pagefault(sb);
 688                 file_update_time(vmf->vma->vm_file);
 689                 down_read(&EXT4_I(inode)->i_mmap_sem);
 690 retry:
 691                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
 692                                                EXT4_DATA_TRANS_BLOCKS(sb));
 693                 if (IS_ERR(handle)) {
 694                         up_read(&EXT4_I(inode)->i_mmap_sem);
 695                         sb_end_pagefault(sb);
 696                         return VM_FAULT_SIGBUS;
 697                 }
 698         } else {
 699                 down_read(&EXT4_I(inode)->i_mmap_sem);
 700         }
 701         result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
 702         if (write) {
 703                 ext4_journal_stop(handle);
 704
 705                 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
 706                     ext4_should_retry_alloc(sb, &retries))
 707                         goto retry;
 708                 /* Handling synchronous page fault? */
 709                 if (result & VM_FAULT_NEEDDSYNC)
 710                         result = dax_finish_sync_fault(vmf, pe_size, pfn);
 711                 up_read(&EXT4_I(inode)->i_mmap_sem);
 712                 sb_end_pagefault(sb);
 713         } else {
 714                 up_read(&EXT4_I(inode)->i_mmap_sem);
 715         }
 716
 717         return result;
 718 }
 719
 720 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
 721 {
 722         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
 723 }
 724
 725 static const struct vm_operations_struct ext4_dax_vm_ops = {
 726         .fault          = ext4_dax_fault,
 727         .huge_fault     = ext4_dax_huge_fault,
 728         .page_mkwrite   = ext4_dax_fault,
 729         .pfn_mkwrite    = ext4_dax_fault,
 730 };
 731 #else
 732 #define ext4_dax_vm_ops ext4_file_vm_ops
 733 #endif
 734
 735 static const struct vm_operations_struct ext4_file_vm_ops = {
 736         .fault          = ext4_filemap_fault,
 737         .map_pages      = filemap_map_pages,
 738         .page_mkwrite   = ext4_page_mkwrite,
 739 };
 740
 741 static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
 742 {
 743         struct inode *inode = file->f_mapping->host;
 744         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 745         struct dax_device *dax_dev = sbi->s_daxdev;
 746
 747         if (unlikely(ext4_forced_shutdown(sbi)))
 748                 return -EIO;
 749
 750         /*
 751          * We don't support synchronous mappings for non-DAX files and
 752          * for DAX files if underneath dax_device is not synchronous.
 753          */
 754         if (!daxdev_mapping_supported(vma, dax_dev))
 755                 return -EOPNOTSUPP;
 756
 757         file_accessed(file);
 758         if (IS_DAX(file_inode(file))) {
 759                 vma->vm_ops = &ext4_dax_vm_ops;
 760                 vma->vm_flags |= VM_HUGEPAGE;
 761         } else {
 762                 vma->vm_ops = &ext4_file_vm_ops;
 763         }
 764         return 0;
 765 }
 766
 767 static int ext4_sample_last_mounted(struct super_block *sb,
 768                                     struct vfsmount *mnt)
 769 {
 770         struct ext4_sb_info *sbi = EXT4_SB(sb);
 771         struct path path;
 772         char buf[64], *cp;
 773         handle_t *handle;
 774         int err;
 775
 776         if (likely(sbi->s_mount_flags & EXT4_MF_MNTDIR_SAMPLED))
 777                 return 0;
 778
 779         if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
 780                 return 0;
 781
 782         sbi->s_mount_flags |= EXT4_MF_MNTDIR_SAMPLED;
 783         /*
 784          * Sample where the filesystem has been mounted and
 785          * store it in the superblock for sysadmin convenience
 786          * when trying to sort through large numbers of block
 787          * devices or filesystem images.
 788          */
 789         memset(buf, 0, sizeof(buf));
 790         path.mnt = mnt;
 791         path.dentry = mnt->mnt_root;
 792         cp = d_path(&path, buf, sizeof(buf));
 793         err = 0;
 794         if (IS_ERR(cp))
 795                 goto out;
 796
 797         handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
 798         err = PTR_ERR(handle);
 799         if (IS_ERR(handle))
 800                 goto out;
 801         BUFFER_TRACE(sbi->s_sbh, "get_write_access");
 802         err = ext4_journal_get_write_access(handle, sbi->s_sbh);
 803         if (err)
 804                 goto out_journal;
 805         strlcpy(sbi->s_es->s_last_mounted, cp,
 806                 sizeof(sbi->s_es->s_last_mounted));
 807         ext4_handle_dirty_super(handle, sb);
 808 out_journal:
 809         ext4_journal_stop(handle);
 810 out:
 811         sb_end_intwrite(sb);
 812         return err;
 813 }
 814
 815 static int ext4_file_open(struct inode * inode, struct file * filp)
 816 {
 817         int ret;
 818
 819         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
 820                 return -EIO;
 821
 822         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
 823         if (ret)
 824                 return ret;
 825
 826         ret = fscrypt_file_open(inode, filp);
 827         if (ret)
 828                 return ret;
 829
 830         ret = fsverity_file_open(inode, filp);
 831         if (ret)
 832                 return ret;
 833
 834         /*
 835          * Set up the jbd2_inode if we are opening the inode for
 836          * writing and the journal is present
 837          */
 838         if (filp->f_mode & FMODE_WRITE) {
 839                 ret = ext4_inode_attach_jinode(inode);
 840                 if (ret < 0)
 841                         return ret;
 842         }
 843
 844         filp->f_mode |= FMODE_NOWAIT;
 845         return dquot_file_open(inode, filp);
 846 }
 847
 848 /*
 849  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
 850  * by calling generic_file_llseek_size() with the appropriate maxbytes
 851  * value for each.
 852  */
 853 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
 854 {
 855         struct inode *inode = file->f_mapping->host;
 856         loff_t maxbytes;
 857
 858         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 859                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
 860         else
 861                 maxbytes = inode->i_sb->s_maxbytes;
 862
 863         switch (whence) {
 864         default:
 865                 return generic_file_llseek_size(file, offset, whence,
 866                                                 maxbytes, i_size_read(inode));
 867         case SEEK_HOLE:
 868                 inode_lock_shared(inode);
 869                 offset = iomap_seek_hole(inode, offset,
 870                                          &ext4_iomap_report_ops);
 871                 inode_unlock_shared(inode);
 872                 break;
 873         case SEEK_DATA:
 874                 inode_lock_shared(inode);
 875                 offset = iomap_seek_data(inode, offset,
 876                                          &ext4_iomap_report_ops);
 877                 inode_unlock_shared(inode);
 878                 break;
 879         }
 880
 881         if (offset < 0)
 882                 return offset;
 883         return vfs_setpos(file, offset, maxbytes);
 884 }
 885
 886 const struct file_operations ext4_file_operations = {
 887         .llseek         = ext4_llseek,
 888         .read_iter      = ext4_file_read_iter,
 889         .write_iter     = ext4_file_write_iter,
 890         .iopoll         = iomap_dio_iopoll,
 891         .unlocked_ioctl = ext4_ioctl,
 892 #ifdef CONFIG_COMPAT
 893         .compat_ioctl   = ext4_compat_ioctl,
 894 #endif
 895         .mmap           = ext4_file_mmap,
 896         .mmap_supported_flags = MAP_SYNC,
 897         .open           = ext4_file_open,
 898         .release        = ext4_release_file,
 899         .fsync          = ext4_sync_file,
 900         .get_unmapped_area = thp_get_unmapped_area,
 901         .splice_read    = generic_file_splice_read,
 902         .splice_write   = iter_file_splice_write,
 903         .fallocate      = ext4_fallocate,
 904 };
 905
 906 const struct inode_operations ext4_file_inode_operations = {
 907         .setattr        = ext4_setattr,
 908         .getattr        = ext4_file_getattr,
 909         .listxattr      = ext4_listxattr,
 910         .get_acl        = ext4_get_acl,
 911         .set_acl        = ext4_set_acl,
 912         .fiemap         = ext4_fiemap,
 913 };
 914